Blog
Hiring an AI agent development company? What they build, what it costs, and how to choose
What an AI agent development company should build for you, what it really costs, the red flags, and 10 questions to ask before you sign.
By Lars ·

Short answer: an AI agent development company designs, builds, tests and maintains software agents that do specific jobs inside your systems. A good one will also tell you when part of the job doesn't need AI at all.
I've been building software for almost 20 years. I started with HTML and CSS for landing pages, and over time ran projects with teams of 30, 40, sometimes 50 people: web and Web3 developers, game developers, designers, UI and UX, and the QA people every serious project needs.
We've used AI in that work for three or four years. It started small: social media and correcting text. Then we used it to check snippets of code. For a long time it couldn't write large pieces of code reliably, and code you know has mistakes in it isn't worth much. If you're holding a sheet you know is wrong somewhere, you might as well have written it yourself.
That has changed. AI can now write pieces of code, test them and help put them together, and the work moves towards integration and operation. But the biggest problem in software hasn't changed, and it decides whether your AI project works: knowing exactly what you're trying to build.
A game or an online shop is fairly straightforward. A complex trading system isn't. Before you build one, someone has to understand the subject, not just the spec in front of them. That comes from experience and from knowing the field. It's true when you build by hand, and it's just as true when you build with AI.
This guide is for buyers comparing AI agent companies. Most of what you'll find for this search is "top 10" lists written by one of the companies on the list. This isn't one of those. It covers what these companies build, what it costs, why projects fail, and the questions to ask before you sign.
What counts as an AI agent (and what doesn't)
| What it does | Example | Use it when | |
|---|---|---|---|
| Chatbot | Answers questions in a conversation | A help widget on your website | People need answers, not actions |
| Automation | Runs fixed steps someone wrote in advance | Every new order is copied to the accounting system | The process is the same every time |
| AI agent | Gets a goal, chooses its own steps and uses your tools to finish the job, with a person approving the steps that matter | Reads a supplier's Word file and photos, creates the products and translates them | The work needs judgement, not just rules |
Many products sold as "agents" are chatbots with a new label. Gartner calls this "agent washing". Of the thousands of vendors claiming agentic AI, it estimates only about 130 offer real agents.
The part most guides leave out is that a working agent is rarely pure AI. Most of what we build is a fixed system first: a database, Python processes that crawl and collect data, and scheduled jobs. AI sits on top, only where judgement is needed. We regularly ask clients: why use AI for this step when a Python script does it, and you only pay for a server that runs for ten minutes? AI usage burns through tokens and budget very quickly.
Small agents, one job each
One agent that does everything doesn't exist, and if it did, it would cost far too much to run. Good agents do one specific thing and are built and tested for that one thing.
Over the years I've built several hundred of these small process agents. Some run on their own, and some we start when we need them. Each one checks a website, reads SEO results, collects market data or analyses a single input. When you need more, you use several small agents and a manager agent. The manager reads what each one found and decides what happens next: flag a security issue, create a task, or do nothing.
On roburnalabs.com you can try a few of ours for free, such as the Idea Validator and the Automation Leak Check.
What AI agent companies actually build
- Product catalogue and content work. For Zaliana, a fashion brand selling in three markets, the admin reads Word files and photos, creates the products and translates them into Spanish and French. (Case study)
- Website, QA and SEO checks. Agents that test a site, read search results and report what changed.
- Research and analysis. Market and trading analysis tools that collect data and pick out what matters.
- Support triage. Read incoming messages, sort them and draft a reply for a person to approve.
- Back-office and order handling. Move information between the tools a business already uses.
- Internal knowledge assistants. Answer staff questions from the company's own documents.
The mistake I made: 200 agents nobody used
I once tried to build an AI team to help run my companies. It reviewed my social media and SEO, came up with suggestions and looked for potential clients. A bit of everything. We coded around 200 agents running in loops.
Technically it worked reasonably well. But the place where my team was supposed to read the results was close to unreadable, and we ended up not using it.
That taught me two things.
- Build agents into a product and a workflow, not the other way round. If you start with agents, keep adding more and then try to build a screen to show what they do, it becomes hard to update, hard to maintain and easy to ignore.
- More reports isn't more value. Imagine a QA bot that checks your website every day. Out of 100 reports, maybe one matters. But you've stopped reading them by then, and you're still paying for the usage. The fix is to connect the agent to a real system. On the day something matters, it becomes a task in your normal task list, with a person who owns it. A Telegram message works for a one-person team. Beyond that, you need to know who saw it, who took action and what happened next.
We rebuilt it the other way round: general bots in Python first, with AI added on top where it helps.
What it costs, and how to think about the price
Published market ranges from agencies in this field:
| What you're buying | Typical range |
|---|---|
| Single-task agent or pilot | $15,000–$50,000 |
| Production agent connected to your systems | $50,000–$150,000 |
| Enterprise programme | $150,000 and up |
| Running costs (AI usage, servers, tools) | $1,000–$10,000 a month |
| Maintenance | 15–25% of the build cost a year |
Sources: SoftTeco, Space-O, Cleveroad, Biz4Group (2026).
In our experience, a small, tightly focused agent can start at around 3,000–5,000. At the top end there's no real limit. It can take a few days or several weeks to build, and a production system can take a few months. The longest part is testing: making sure the output is what you expect, cutting the AI usage down, and getting the whole thing properly structured.
AI makes some things cheaper and shortens delivery. It hasn't made good development cheap. Experts who understand both the technology and your field cost more as they gain experience, and over the years many of our own developers have been hired away by data centres and large B2B platforms that could pay more.
A better way to judge the price is value. Say you pay a video editor somewhere between $500 and $30,000, depending on where they work and what they do. You replace most of that work with an automated process using tools like Claude, ElevenLabs and a video model, for $100–$200 a month. If you save $2,000 a month, that's $24,000 in a year. At that point, whether the build cost $5,000 or $20,000 matters less than three other things: it works, it's reliable, and it can be improved as your work changes.
The real price of software is what it costs you when it fails. If it breaks, your customers feel it, and the longer it's down, the more revenue you lose.
Why agent projects fail
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of cost and unclear value. In LangChain's 2026 survey, quality is the biggest barrier to getting agents into production. From our side of the table, failures come back to the same few causes.
- The scope wasn't really shared. The client and the developer agreed on a document but pictured different things. When changes arrive in the middle of a project, or worse at the end, it doesn't matter whether the price was 3,000 or 50,000: part of it has to be rebuilt. Late changes eat the testing budget, which is one of the biggest real costs in any build.
- Nobody caught what was wrong. AI can tell you something works when it doesn't. You can't ask it to fix a problem it doesn't know is there. Until a person understands the issue, nobody can fix it. QA is still the area where AI fails most.
- It works, but nobody can operate it. A system can match the scope perfectly and still be worthless if the client's team can't use it. Non-technical clients often need it built differently.
Red flags when choosing a company
- They can't explain the difference between a chatbot, an automation and an agent.
- They have no testing plan, or the only proof that it works is that the AI said so.
- There are no human approval points.
- They don't estimate the monthly running cost.
- They have no maintenance plan.
- They never ask about your user flows, your onboarding or who will operate it day to day.
- On a rescue, they want to rebuild everything from scratch. "This code is bad" often means "I would have done it differently." Different developers structure code very differently, with or without AI. A good company can work on top of someone else's code when that's the best solution.
Access and safety: what must be true before an agent touches real systems
An influencer account might be happy to let agents publish a thousand posts a day. A company has a brand to protect, details that must be correct and, often, rules it has to follow.
- A person stays in the loop for anything that represents your company, moves money or reaches customers. AI makes mistakes and misreads context. People do too, which is why there's a review step.
- Sensitive things are out of reach. If you hold client data, work in finance or have strict security, the agent must not be able to see or touch certain systems at all. In some jurisdictions you aren't even allowed to let it.
- Access is narrow and short-lived. Agents get only the access the job needs, read-only where possible, and it expires.
Maintenance is part of the price
Software that runs needs looking after. Security issues appear in widely used tools: in December 2025 a critical flaw in React Server Components (CVE-2025-55182) forced emergency updates on sites everywhere. Access tokens expire and must be renewed on the servers. Hosting providers patch their servers, sometimes at very short notice, and a process stops.
There's a second reason too. Almost everything I've built, the moment it went live and real users touched it, we found ways to make it better and save even more time. A good maintenance agreement covers both: priority support, where someone can drop what they're doing when it breaks, and ongoing care for updates and improvements.
10 questions to ask before you sign
- Can you walk me through what it will and won't do, the user flow, and how my team or clients get onboarded?
- Will you record our calls and check the final requirement sheet against them?
- Which parts need AI, and which can be ordinary code?
- How do you test it, and who checks the results apart from the AI?
- What happens when it's wrong? Where are the human approval points?
- What will it cost to run each month?
- What can the agent see and touch, and what is kept out of its reach?
- Who owns the code and the data, and how are deletion requests handled?
- How are scope changes handled, and what do they cost?
- What does maintenance cover, and how fast do you respond when something breaks?
The one thing I wish every buyer knew
Whether you hire an AI company, an agency or a freelancer, the outcome depends on one thing: the client and the developers understanding the same result.
Talk through the process, not just the features. How does a user move through it? How do you onboard your clients onto the platform, and what do you need for that? If you ask the right questions, you get the right answers, but those answers often never make it into the requirement sheet. Onboarding people is hard, because behaviour changes from today to tomorrow while the code stays fixed. Sometimes the development is the easy part.
This is also the difference between a freelancer and an agency. A good agency has people who know different industries, and someone who understands your field. If nobody understands what you're building, you can't build it well, with AI or by hand.
Record your calls, then check the requirement sheet against them
This is the most practical advice I can give, to clients and to other developers and agencies alike: record every conversation about the project.
On a bigger project, the details get discussed over several meetings. Before you sign off the final requirement sheet, even if you've already paid a deposit, go back through the recordings and check every detail against the document:
- What did we agree the user flows are?
- Which small but important functions did we talk about?
- What's the outcome, and what reporting comes with it?
- What does the admin panel need to do?
- When a user in the EU asks for their data to be deleted, how is that handled? Other countries and other industries bring their own data rules, and the system has to follow them.
Make sure each detail was discussed clearly, recorded clearly and written into the document you both agree on. Always make a document, however big or small the job. It's a boring step, and developers don't like doing it, but it protects both sides. Think of it like a physical product: where the logo goes and which colours to use matter, and a digital product has just as many details. They're just different ones.
I record several hours of meetings every day with developers and clients. Some days we record 15 or 16 hours. We review them against the requirements and, where we can, note exactly where in the conversation each decision was made. Then the developer can explain each point in a way the client agrees with.
If you're a client, make sure whoever runs your project does this, whether that's a project manager, an agency or a freelancer. I'd recommend always working through a manager who isn't the one writing the code and can sit between you and the developers. That's the smoothest way to run a project.
In all my years, I wouldn't say failed projects come from developers making mistakes. Developers want to deliver a good product, and clients want a good product. It comes down to misunderstandings. This applies far beyond software: in any custom deal, B2B especially, record the conversation, make sure both sides understand the terms, and write them down.
How we work
We spend time with you up front, so we both understand what will be built and how it will be used. We build the fixed system first and add AI only where it earns its place. Every part is tested by people and by a test suite written separately from the code. There are human approval points wherever your brand, money or data is involved, and a maintenance plan once it's live.
Book a 30-minute scoping call or, if you're not sure you need an agent at all, run the free Automation Leak Check first.
FAQ
Will an AI agent replace my staff? For most companies, no. It takes over a specific, repetitive part of the work, and people review what matters. Treat it as a helping tool added to a workflow.
Should I hire a freelancer or an agency? A freelancer can be right for a small, well-defined job. For anything touching several systems or a field with its own rules, an agency brings people across disciplines and continuity when someone leaves.
How long until it pays for itself? Work out what the task costs you today per month, subtract the running costs, and divide the build price by what's left. If you save $2,000 a month, a $10,000 build pays back in about five months.
Can you take over an agent someone else built? Usually, yes. Rebuilding from scratch isn't the default. We first check whether the existing work is sound, even if it's structured differently from how we'd do it.
Written by Lars, founder of Roburna Labs · Nov 10, 2026
Builds companies, and keeps looking for a better way to run them. Currently in Peru.